基于加权特征空间信息视觉词典的图像检索模型

Image retrieval model based on visual vocabulary with weighted feature space information

针对传统的视觉词袋模型中视觉词典对底层特征量化时容易引入量化误差,以及视觉单词的适用性不足等问题,提出了基于加权特征空间信息视觉词典的图像检索模型。从产生视觉词典的常用聚类算法入手,分析和探讨了聚类算法的特点,考虑聚类过程中特征空间的特征分布统计信息,通过实验对不同的加权方式进行对比,得出效果较好的均值加权方案,据此对视觉单词的重要程度加权,提高视觉词典的描述能力。对比实验表明,在ImageNet图像数据集上,相对于同源视觉词典,非同源视觉词典对视觉空间的划分影响较小,且基于加权特征空间信息视觉词典在大数据集上更加有效。

Concerning the quantization error when the local features were quantified by the visual vocabulary in traditional Bag-of-Visual-Word （BoVW） model,an image retrieval model based on visual vocabulary with weighted feature space information was proposed.Considered the clustering method which was used to generate the visual codebook,the statistic information of the feature space was analyzed during the clustering process.Through the comparison of different weighting methods by experiments,the best weighting method,mean weighted average,was found to weight the visual words to improve the descriptive ability of the codebook.The experiment on ImageNet dataset shows that,compared to homologous visual codebook,non-homologous visual codebook has less impact on dividing the visual space,and the effects of the weighted feature space based visual codebook on big dataset are better.